DiffSegNet: 3D Point Cloud Instance Segmentation Model Based on Differential 3D U-Net

Table 1. Category statistics
View table
Table 1. Category statistics
Data Total Train Validation Test
Transformer 350 245 70 35
Insulator 1340 938 268 134
Electric wire 720 504 144 72
Bushing 1220 854 244 122
Pole 400 280 80 40
Total data 4030 2821 806 403

Table 2. Real label
View table
Table 2. Real label
Category Transformer Insulator Electric wire Bushing Pole
Real label
Mask color

Table 3. Comparative experimental results of various models on self-built dataset

View table

Table 3. Comparative experimental results of various models on self-built dataset

Model	AP₅₀ /%					mAP₅₀ /%
Model	Transformer	Electric wire	Pole	Insulator	Bushing	mAP₅₀ /%
PointGroup^［26］	65.72	56.41	64.16	55.64	58.45	60.07
DyCo3D^［27］	64.32	60.31	65.25	54.67	59.27	60.76
HAIS^［28］	69.73	63.46	64.14	56.80	60.33	62.89
SoftGroup^［14］	70.44	64.02	65.12	57.14	61.09	63.55
Proposed	73.31	67.24	67.93	60.57	63.87	66.58

Table 4. Comparative experimental results of various models on ScanNet dataset

View table

Table 4. Comparative experimental results of various models on ScanNet dataset

Category	AP₅₀ /%
Category	PointGroup	DyCo3D	HAIS	SoftGroup	Proposed
Bathtub	84.12	80.41	80.47	82.61	84.45
Bed	76.65	83.42	82.71	80.32	81.51
Shelf	61.24	83.28	84.45	81.36	83.52
Cabinet	50.44	55.14	56.48	61.42	63.67
Chair	79.41	78.46	79.42	83.88	84.92
Curtain	69.45	66.13	68.41	72.71	71.80
Desk	38.13	46.28	48.49	56.34	64.52
Counter	11.23	15.49	24.32	24.12	28.45
Door	44.06	43.80	45.45	56.74	61.05
Picture	46.41	48.32	54.32	60.91	67.31
Fridge	59.38	55.63	57.28	69.32	72.15
Sink	66.15	56.18	67.31	65.44	70.62
Sofa	75.67	75.43	76.44	75.72	78.46
Table	58.15	65.85	67.13	67.58	70.92
Toilet	83.13	82.34	80.23	82.14	83.24
Window	53.43	55.31	56.49	56.04	55.47
Other	57.15	57.26	58.12	65.65	72.94
mAP方正汇总行₅₀ /%	59.66	61.59	63.97	67.20	70.29

Table 5. Results of ablation experiment 1

View table

Table 5. Results of ablation experiment 1

Serial number	VVEM	DFFM	DFM	AP₅₀ /%					mAP₅₀ /%
Serial number	VVEM	DFFM	DFM	Transformer	Electric wire	Pole	Insulator	Bushing	mAP₅₀ /%
1	—	—	—	70.43	64.02	65.12	57.14	59.32	63.20
2	√	—	—	71.52	65.64	66.75	58.42	60.87	64.64
3	√	√	—	72.74	66.76	67.29	59.74	62.09	65.72
4	√	√	√	73.31	67.24	67.93	60.57	63.87	66.58

Table 6. Results of ablation experiment 2
View table
Table 6. Results of ablation experiment 2
Serial number Regularization AP₅₀ /% mAP₅₀ /%
Transformer Electric wire Pole Insulator Bushing
1 — 68.73 63.91 64.67 56.55 58.43 62.46
2 √ 73.31 67.24 67.93 60.57 63.87 66.58

Tools

Get Citation

Copy Citation Text

Gang Yang, Na Zhang, Fan Hu, Hua Yu. DiffSegNet: 3D Point Cloud Instance Segmentation Model Based on Differential 3D U-Net[J]. Laser & Optoelectronics Progress, 2025, 62(8): 0815003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites